PLOTSHOT: Generating Discourse-constrained Stories around Photos

Page created by Howard Wagner
 
CONTINUE READING
PLOTSHOT: Generating Discourse-constrained Stories around Photos
P LOT S HOT: Generating Discourse-constrained Stories around Photos

                                     Rogelio E. Cardona-Rivera1,2 and Boyang Li2
                  1
                      Liquid Narrative Group, Department of Computer Science, North Carolina State University
                                                       2
                                                         Disney Research
                                         recardon@ncsu.edu, albert.li@disneyresearch.com

                            Abstract                                      We propose one such story generator, which we
                                                                    call P LOT S HOT. This system considers optional discourse
  Story generators typically adopt a pipelined model of
                                                                    goals, which represent available discourse materials, and
  generation wherein fabula structure is decided independently
  and prior to discourse structure. In this paper, we propose       attempts to incorporate them into the generated story.
  a novel story generator, P LOT S HOT, capable of reasoning        Our formulation combines classical narrative planning with
  over discourse materials during fabula generation such that       oversubscription planning (Smith 2004). The former affords
  these materials meaningfully constrain the development of         reasoning about authorial intent, i.e. achieving specific
  a causally and intentionally coherent story. P LOT S HOT in-      story outcomes and intermediate states. The latter affords
  corporates user-supplied photographs as optional story states     reasoning about optional intermediate states and a hard cost
  through an oversubscription planning paradigm. Further,           budget to fit stories to fixed-length formats, e.g. feature-
  to leverage existing work on planning-based models of             length movies or booklets. To solve the hybrid problem, we
  generation, we present a technique to compile the photo           present a technique to select amongst available discourse
  story planning problem to classical narrative planning. Our
                                                                    materials and compile the novel problem into a classical
  system attempts to maximize quality of an illustrated story
  by analyzing the affinity between a photo and the action it is    problem in order to leverage existing work in planning-
  meant to depict. An evaluation of generated artifacts shows       based narrative generation.
  advantage over heuristic baseline techniques.                           The specific domain of application for this paper is
                                                                    generating illustrated stories around a set of user-supplied
                                                                    photos. We represent photos via scene graphs, semantic
                         Introduction                               networks that describe the photos’ content. As a photo
Research in story generation has advanced a bipartite               depicts the current world state, it can be displayed only
representation of narratives (Young 2007). One part is              when all of its content has been established. Since there are
termed the fabula, which is the conceptualization of the            often multiple locations where a photo may be displayed,
story world including actants that exist and events that            we present a photo placement strategy based on the affinity
transpire. The other part is termed the discourse, which            between photos and story actions. Our evaluation of gen-
includes elements responsible for the story’s telling. Most         erated photo booklets demonstrates the effectiveness of the
existing story generation work (for reviews see Gervás 2009;       affinity-based strategy over heuristic baseline techniques.
Ronfard and Szilas 2014) has focused on a fabula-driven
pipeline. This approach commits to a fabula by simulating           Contributions In this paper(1) we propose a novel chal-
the story world as a sequence of events, and then creates a         lenge for selective incorporation of discourse materials into
discourse by selectively presenting those events; fabula is         fabula generation, which is beyond the capabilities of state-
generated independently of and prior to the discourse.              of-the-art narrative planning systems; (2) we formulate the
      However, to create captivating fabula, it is important        challenge as a hybrid classical/oversubscription planning
to consider discursive constraints. For instance, a movie           problem and present a technique to compile the hybrid
production may choose to include a car chase scene because          problem into a classical problem; (3) we define a metric
of its visual appeal. As a result, the fabula must adapt            to evaluate the affinity between a photo and the story event
and provide causal and motivational justification for the car       it depicts, and (4) demonstrate that an affinity-based photo
chase in order to maintain story coherence. This process can        placement strategy is significantly better than baselines.
be understood as a discourse placing constraints on a fabula.
Further, there may be multiple cinematic scenes that have                                Related Work
different visual appeal and require different story structures.
                                                                    Fabula-driven approaches have been the dominant model
A story generator should select visual scenes to optimize for
                                                                    in automated story generation (e.g. Aylett, Dias, and
and balance visual appeal with story coherence.
                                                                    Paiva 2006; Li 2015; Teutenberg and Porteous 2015; Barot,
Copyright c 2016, Association for the Advancement of Artificial     Potts, and Young 2015). However, there are a few exceptions
Intelligence (www.aaai.org). All rights reserved.                   where discourse reasoning substantially impacts fabula.
PLOTSHOT: Generating Discourse-constrained Stories around Photos
standing
                                                                                                            girl
                                                                                                                           blonde

                                                                                                               wearing

                                                                                                                   dress        flowery

                                                                                                     holding
                                                                                                                       red
                                                                                                     balloon
                                                                                                                       pink

                                                                       Binary Literals           Unary Literals      flowery(?dress)
Figure 1: The P LOT S HOT system architecture. Boxes with             wearing(?girl, ?dress)     standing(?girl)     red(?balloon)
text represent algorithmic modules.                                   holding(?girl, ?balloon)   blonde(?girl)       pink(?balloon)

                                                                    Figure 2: An example of a scene graph and its grounding
      The Slant story generation system (Montfort et al.            parsed from an image (top) and the extracted logical literals
2013) detects aspects of the verbs during discourse gen-            (bottom). The scene graph contains objects (in yellow),
eration and determines the fabula’s match to different              relations (in blue) and attributes (in green).
genres. After a specific genre has been identified, additional
constraints are posed to the fabula generator. Unlike their
work, our system works with discourse constraints supplied          narrative planning, a Photo Placement module places photos
by users in a novel planning formulation. Piacenza et               in the story. The process employs a measure of compatibility
al. (2011)’s video recombination system is very similar to          between story actions and photos, defined as photo-action
our work. They learn some basic semantic units of video and         affinity. We describe these modules in following sections.
a Markov transition model between the units. All actions
in the fabula plan must have a corresponding video clip.            Representing Photos
The planner sends requests for video clip. If a request is          Computationally parsing the semantic meaning of images
denied, the story can be replanned. In comparison, we allow         is an open and difficult challenge. Recently, Krishna et
actions without photos and describe such actions with text.         al. (2016) proposed parsing images as scene graphs. A scene
In our formulation, discourse materials are optional goals          graph is a directed graph consisting of objects (e.g., people,
that the planner explicitly attempts to achieve, not requested      places, things), relationships between objects (e.g., on-top-
after a plan is created. Radiano et al. (in revision) select and    of, wearing), and attributes of objects (e.g., color, shape).
fit photos to a given story template instead of dynamically         Each object is grounded through a bounding box.
adapting stories to suit photos.                                          We assume that we have accurately parsed scene
      Some interactive story systems aim to satisfy manda-          graphs from user-supplied photos, and that these graphs
tory intermediate story goals (Riedl 2009); these are similar       preserve identities for actors across the photos using tech-
to ours as they all need to work with events that must happen       niques like that of Dai et al. (2015). Such a parser is beyond
or must have happened, given discourse events that have             the scope of this work. The main reason for adopting this
been observed. Robertson and Young (2014) propose an                representation is its active use in computer vision research.
algorithm for modifying early events that are unknown to the
                                                                          To bridge with the representation used by narrative
player but are causally related to later, known events. These
                                                                    planners, we straightforwardly convert a scene graph to a set
early events can be replaced as long as preconditions of
                                                                    of logic literals: relations can be converted to binary literals
the known events can be satisfied. Tomai (2014) developed
                                                                    and attributes to unary literals. We opt to treat objects in the
a system that is free to assert events that may reasonably
                                                                    extracted literals as variables, so that they can be bound with
have happened at the same time or slightly before observed
                                                                    planning problem objects. Figure 2 shows an example of an
player actions. In addition to mandatory goals, our work also
                                                                    example scene graph and logic literals extracted from it.
supports optional intermediate goals that may be abandoned
if they are deemed too costly.                                            There is a potential for domain disparity between the
                                                                    scene graphs and our planning domain. That is, some literals
                                                                    in the photos (e.g., hair color as in Figure 2) may not be
                The P LOT S HOT System                              specified anywhere in the planning domain definition. As a
P LOT S HOT is a narrative generation system for photo-             result, they can never be achieved. To get around this issue,
illustrated stories. Its architecture is illustrated in Figure 1.   we remove from the parsed literals the predicates not defined
P LOT S HOT relies on a Photo Parser module to parse a photo        in the planning domain. A photo with no literals left cannot
into a set of logic literals, which form an optional discourse      be used in the story.
goal. A photo story planning problem contains optional
discourse goals and a planning domain designed by domain
                                                                    Photos as Optional Goals in Planning
engineers. The ILP Compilation module selects optional
discourse goals to create an approximate classical problem          In this section, we formulate the photo-story problem based
that can be solved by an off-the-shelf narrative planner. After     on the needs for story generation from photos.
PLOTSHOT: Generating Discourse-constrained Stories around Photos
Background Young et al. (2013) model a story as a plan,               Go , and a set of mandatory intermediate goals GM . Each
or a sequence of actions that transforms an initial state s0 to       goal g ∈ Go ∪ GM ∪ {g∞ } is a set of logic literals. An
achieve a set of goal literals g∞ . Each action a is associated       intermediate goal g i is achieved by an action sequence p if
with a set of preconditions PRE(a), effects EFF(a), and a cost        any state in τ (s0 , p) contains all literals in g i . The final story
c(a). Effects are bipartite: positive literals are recorded in an     goal g∞ is achieved by p if the last state in τ (s0 , p) contains
add list, and negative literals are recorded in a delete list. An     all literals in g∞ .
action is applicable in a state when its preconditions are true              The following notations define plan quality. The
in that state. After an action is executed, literals in the delete    function c(a) measures the cost of action a. Cmax is the
list are removed from the state and literals in the add list are      limit on total cost. A benefit value b(g o ) is defined for each
added to it. In classical planning, goal satisfaction is a hard       optional intermediate goal g o ∈ Go , which represents the
constraint: a plan is valid only if all goal literals are made        aesthetic quality of the photo. In this paper, we assume
true. Conversely, cost is a soft constraint being optimized.          aesthetic quality of a single photo is given to us as a single
       Over-subscription planning (OSP) (Smith 2004) is a             number, as much research has been done on its automatic
different formulation. An OSP problem contains a set of               assessment (e.g., Joshi et al. 2011; Morgens and Jhala 2013).
optional goals and a maximum cost Cmax . A valid solution             In later sections, we consider another aesthetic concern of
must cost less than Cmax . An example OSP problem is for a            the connection between photos and actions. A valid solution
robot to collect rewards from different locations subject to a        to the problem is a sequence of actions p that can be applied
finite amount of fuel. Different locations can have different         in the initial state, achieves all of GM and g∞ , and has a total
benefits and different routes consume different amounts                     Pk
                                                                      cost i=1 c(ai ) ≤ Cmax . Among valid solutions, we seek
of fuel. We want to maximize the sum of benefits before               one that maximizes the sum of the benefits of all achieved
exhausting the fuel. Unlike classical planning, in OSP goal           optional goals.
satisfaction is a soft constraint and cost is a hard constraint.             Domshlak and Mirkis (2015) argue that OSP cannot
The Use Case We formulate the photo story planning                    reduce to classical planning without incurring a loss of plan
problem as a hybrid of OSP and classical planning with both           quality. Mixing optional and mandatory goals further adds to
optional and mandatory goals. This is driven by our use case          the difficulty of the problem. However, compiling our photo
of story generation around user-supplied photos.                      story planning problem to classical narrative planning allows
      Photos in our stories are naturally optional because            us to benefit from decades of extensive research on the latter
users usually have many more photos than we can put into              problem. In the next section we present an approximate
a story, so some selection of photos is necessary. A photo is         compilation technique.
considered to depict a state of the underlying story world.
To maintain story coherence, all facts that are present in a          The Compilation: Choosing Mandatory Goals
photo must be causally established (but not always explicitly         from Optional Goals
presented) before the photo can be shown.
                                                                      We propose a technique for selecting some optional goals
      Mandatory goals arise from the need to specify                  and convert them to mandatory goals in classical planning.
story outcomes and the need to represent authorial inten-             This is similar to Smith (2004) but the difference is that we
tions (Riedl 2009), which indicate intermediate story states          consider both optional and mandatory goals in the selection.
through which all valid stories must pass. For example, in
                                                                            Goal selection can be thought of as traversing a
the story world of Cinderella, an intermediate story state
                                                                      directed graph where each node represents a goal. Each
could be that Cinderella leaves her shoe at the palace at some
                                                                      directed edge has a weight, representing the additional cost
point. One condition for the story outcome is she gets her
                                                                      needed to reach the second goal after achieving the first
shoe back. The ability to specify authorial intentions and
                                                                      goal. A valid path starts at the initial state, ends at the
outcomes grants effective control over the narrative arc to
                                                                      story goal, and visits every mandatory intermediate goal and
the narrative domain engineer. A upper bound on cost allows
                                                                      some optional intermediate goal. We seek a valid path that
us to fit a story into a fixed-length format such as a poster.
                                                                      maximizes the total benefit of visited optional goals. This
The Photo Story Planning Problem Formally, a photo                    formulation is similar to the traveling salesman problem and
story planning problem is a tuple hS, s0 , g∞ , Go , GM , A,          its variant, the orienteering problem, but also different due to
σ(·), c(·), Cmax , b(·)i. We first explain notations related          the existence of both optional and mandatory components.
to state trajectories. S is the set of possible states. s0 ∈          After we find the optimal path, all optional goals on the path
S is the initial state. A is a set of possible actions. An            are turned to mandatory goals in the planning stage.
action a ∈ A can be applied to a state s ∈ S if its                         We use integer linear programming (ILP) to solve this
preconditions PRE(a) ∈ s. σ(·) denotes the deterministic              problem. Although ILP is NP-complete, several fast off-the-
transition function; we let s0 = σ(s, a) denote that applying         shelf solvers exist. For the ILP problem, we introduce vertex
a in s results in a new state s0 . For a given action sequence        variables Vi ∈ {0, 1}, ∀i ∈ {1, . . . , n} and edge variables
p = [a1 , . . . , an ], we let τ (s, p) denote the state trajectory   Eij ∈ {0, 1}, ∀i, j ∈ {1, . . . , n}. The vertex variables
resulted from applying p sequentially to s. That is, τ (s, p) =       include: the initial state V0 , the story goal Vn , k optional
[σ(s, a1 ), σ(σ(s, a1 ), a2 ), . . .].                                intermediate goals V1 , . . . , Vk , and n − 1 − k mandatory
      The following notations define plan goals. Goals in the         intermediate goals Vk+1 , . . . , Vn−1 . Edge variables Eij and
problem include the story goal g∞ , a set of optional goals           Eji represent two oppositely directed edges between every
vertex pair Vi and Vj , with the exception that V0 does                    Photos                Story Plan           Placement
not have incoming edges and Vn does not have outgoing
                                                                       in(hansel,cage)       travel(hansel,house)
edges. Setting an edge variable or a vertex variable to 1
puts the edge or the vertex on the optimal path. The edge             at(gretel,house)
weight cij reflects the cost incurred by going from Vi to Vj .                              lock-up(witch,hansel)
                                                                      at(witch,house)
If the ith vertex is optional, its benefits bi is greater than
                                                                      at(hansel,house)       travel(gretel,house)
zero; otherwise bi = 0. The ILP maximizes the following                at(witch,house)
objective function:                                                      free(hansel)
                              X                                                             call-for(hansel,gretel)
                         max      bi V i
                                i                                                            defeat(gretel,witch)
while respecting the following constraints:
                                                                                             unlock(gretel,hansel)
  P P
•    i   j cij Eij ≤ Cmax : the total cost of the plan is no
  more than a predefined maximum.
• Vi = 1, ∀i ∈ {1, k + 1, . . . , n}: the initial state, the goal   Figure 3: Additional decisions are needed for placing
  state, and all authorial goals must be on the path.               photos. For the three photos on the left, the three bars on
  P                                                                 the right show their possible placement in the story. As an
•    j Eji = Vi , ∀i 6= 1: every selected vertex has an             illustration, we manually picked the best locations for these
  incoming edge, except the initial state.                          photos, as indicated by the bright-colored segments.
  P
•    j Eij = Vi , ∀i 6= n: every selected vertex has an
  outgoing edge, except the goal state.                               state s0 to the ith intermediate goal. Following Garcı́a-
We prevent cycles by introducing auxiliary variables                  Olaya, de la Rosa, and Borrajo (2011), we compute the
Ui , ∀i ≥ 2 and the following constraints (Miller, Tucker,            state at Vi by applying the relaxed plan’s actions in
and Zemlin 1960):                                                     sequence, applying both add lists and delete lists and
                                                                      ignoring preconditions. We denote the state obtained from
• 2 ≤ Ui ≤ n, ∀i ∈ {2, . . . , n}
                                                                      applying R0i as sR  0i . We then compute a relaxed plan from
• Ui − Uj + (n − 1)Eij ≤ (n − 2), ∀Ui , Uj , i 6= j                   sR  to the j th
                                                                                      intermediate   goal and estimate the distance
                                                                       0i
Finally, we account for mutual exclusions between goals. It           cij accordingly.
is likely that users provide multiple photos depicting similar      3. The distance cin from an intermediate node Vi to the goal
underlying content, such as photos of the same subjects in             node Vn is computed with the relaxed plan from sR0i to the
slightly different poses or angles. The literals from the scene        story goal g∞ .
graphs of the similar photos will be equivalent. To avoid
repetition, we only allow one such photo in our story. This is            We use the ILP to compile the photo story problem to
achieved by the mutex conditions. A mutex set M contains a          a classical planning problem. All optional discourse goals
number of vertices such that only one in M can be selected          that are visited by the optimal path become mandatory
in the path. That is,                                               intermediate goals. Unvisited discourse goals are discarded.
   P                                                                After the compilation, the problem can be solved by any off-
•     i Vi ≤ 1, ∀i ∈ M                                              the-shelf narrative planner. If the optimal plan found exceeds
Unlike Smith (2004) who estimated edge costs based on               the cost limit Cmax , the optional discourse goal ordered
a manual sensitivity analysis of plan cost, we estimate the         last in the plan is removed and planning is attempted again.
edge cost automatically via the use of relaxed plans. A             In this paper, we use the Glaive narrative planner (Ware
relaxed plan is a directed layered graph that is a subgraph         and Young 2014), in which all actions performed by story
of a planning graph (Blum and Furst 1997). A relaxed                characters must serve a character intention but not all
plan starts in state s and reaches s0 by ignoring the delete        character intentions succeed.
lists of all actions; the state contains a monotonically non-
decreasing set of literals as more actions are applied to           Photo Placement
it. The length of a relaxed plan that solves a planning             Photos in a high-quality story should be aesthetically
problem represents an estimate of the actual cost to solve          pleasing by themselves and also exhibit concord with the
said problem. In our case, we use relaxed plan to estimate          story. As discussed earlier, we capture aesthetic quality of
the distance cij between the nodes of our graph. We consider        individual photos as benefits of optional goals in the ILP
three cases:                                                        problem. We now consider the compatibility between photos
                                                                    and the story, which we use to place photos.
1. The distance from V0 to an intermediate node Vi is                     A photo may be placed wherever all of its literals
   denoted by c0i and is computed with the relaxed plan from        are established, which leaves significant freedom for its
   the initial state s0 to the ith intermediate goal.               ultimate position. Figure 3 illustrates one such example,
2. The distance cij from an intermediate node Vi to                 which considers three photos with different semantic con-
   another intermediate node Vj is computed as follows.             tent. A story plan is shown in the middle, where each
   We first compute a relaxed plan R0i from the initial             box represents one action. The three colored bars on
the right show possible positions for the three photos             The family went from the Campsite
respectively. For example, the blue photo can be shown right       through the Base, all the way to the
after the action travel(hansel,house) or right after               Mountain. At the Mountain, they
unlock(gretel,hansel). The bright small boxes                      boarded the minecart. Then, they rode
show the ideal positions for the three photos.                     the minecart from the Mountain to the
                                                                   Mountaintop. From the minecart, the
Photo-Action Affinity In order to place photos relative to         family distracted the Yeti with their
actions, we develop the metric photo-action affinity based on      screams.
structural properties of the story plan, which we believe to
capture how well a photo depicts a story action. The affinity
can be positive or negative. In general, a negative affinity
indicates the photo and the action should not be placed
together. Our metric depends on four features that relate an
action a and a discourse goal g o :
                                EFF(a) ∩ g o
          Support(a, g o ) =
                                     go                                                                    Page 3 of 5

                                                         o
                                {l | l ∈ EFF(a) ∧ ¬l ∈ g }         Figure 4: One page in the story booklet generated by the
          Conflict(a, g o ) =                                      A FFINITY strategy.
                                             go
                                PRE (a) ∩ g o                             Table 1: Weights used for the A FFINITY strategy.
         Synergy(a, g o ) =
                                PRE (a) ∪ g o
                                                                           β1      0.25       β2      -0.5               β3   5.0   β4   3.0
                       o        CHAR (a) ∩ CHAR (g o )
        CharSym(a, g ) =                                           all scores by the absolute value of the largest negative entry,
                                CHAR (a) ∪ CHAR (g o )
                                                                   and shift them back after running the algorithm.
where C HAR(·) denotes the set of story characters in-                   If, in the returned result, a photo is assigned to an
volved in an action or a set of literals. In plain English,        action with a negative affinity, we remove that photo from
Support(a, g o ) measures how many of g o ’s literals are estab-   the generated story. Because of this, some photos may be
lished by a. Conflict(a, g o ) measures opposing conditions a      excluded; we require that every shown photo is accompanied
sets up against g o , which must then be toggled for achieving     by one action such that it has some accompanying text. It is
g o . Synergy(a, g o ) measures common preconditions of a          possible to avoid this issue by tweaking the ILP formulation
and literals in g o , as identical conditions can be achieved      and letting ILP select photo-action pairs instead of only
simultaneously. CharSym measures story characters shared           photos, but we defer that for future work. In this paper,
between the action and the photo of g o . We define photo-         we focus on the affinity-based photo placement strategy and
action affinity as the linear combination of the four features:    evaluate it in the next section.
  Affinity(a, g o ) = β1 Support(a, g o ) + β2 Conflict(a, g o )
                    + β3 Synergy(a, g o ) + β4 CharSym(a, g o )                                    Evaluation
where β1 , β2 , β3 , and β4 are linear weights. For the            In this section, we evaluate the photo placement strategy
experiment in this paper, we tuned the weights manually,           based on the affinity measure with a user study.
but learning these weights as a linear regression is also          Methodology For comparison, we created two heuristics
straightforward.                                                   for photo placement. The R ANDOM strategy places a
                                                                   photo at a random legal location. The E ARLY strategy
Optimizing Placement To place photos using the photo-
                                                                   places a photo at the earliest legal location.2 Weights used
action affinity metric, we treat photo placement as an
                                                                   for computing photo-action affinity in our affinity-based
assignment problem: each photo can be assigned to one
                                                                   strategy (A FFINITY) are shown in Table 1.
action in the story plan, and we seek the assignment that
                                                                         All photo placement strategies started with the same
maximizes overall affinity. After a narrative plan is returned
                                                                   story plan that involved 16 actions in a Disney-themed
by the story planner, we set up an n × m matrix M for
                                                                   domain and 5 photos. One story was created from each
n actions and m discourse goals, where the element Mij
                                                                   placement strategy. Each story was presented to participants
represents the photo-action affinity between the j-th photo
                                                                   as a photo booklet with landscape letter-sized pages. In each
and the i-th action. If the j-th photo does not have its
                                                                   booklet, one page contained a maximum of one photo. A
literals true after execution of the i-th action, we set that
                                                                   photo was always placed on the same page with the action
photo-action affinity score to −1. After setup, we used the
                                                                   it was paired with. Actions were translated into text using
Hungarian algorithm to solve the assignment problem in
                                                                   rule-based templates. Figure 4 shows one page from the
O(n3 ) time.1 This algorithm requires that all matrix entries
are non-negative. Before running the algorithm, we increase           2
                                                                        We experimented with a L ATE heuristic, which always places
                                                                   the photo at the last legal position. Since this heuristic works poorly
   1
       Assuming that n > m. Otherwise, it runs in O(m3 ) time.     in practice, we excluded it from this study.
A FFINITY story. Every photo had at least 3 possible slots               Table 2: Results from the ranking of three stories.
(avg = 4.75) where it could be placed, except for one photo
that was always at the end.                                                 Strategy                      Votes      p-value
      We recruited 21 participants, with an age range from                  A FFINITY > E ARLY            16/21      0.013
21 to 40, all of whom have at least a bachelor’s degree. Each               A FFINITY > R ANDOM           20/21      R ANDOM              21/21      M dE ARLY > M dR ANDOM ,        beat non-informed baseline techniques. All told, we have
where M d denotes the median rating. In all 9 tests, we reject      demonstrated that complex decision making is needed to
the null in favor of our alternate at the p = 0.05 level,           ensure the coherence between fabula and discourse materials
indicating that the A FFINITY story is perceived to have the        during story generation.
best photo positioning and is the most coherent.                          As a first step toward selective utilization of discourse
                                                                    constraints during fabula generation, we believe this paper
Discussion Although we expected the A FFINITY-based                 opens new research avenues. One such avenue is the
approach to comfortably beat the R ANDOM baseline, the              development of a unified fabula/discourse search paradigm,
comparison between A FFINITY and E ARLY was initially               wherein discourse can constrain fabula during generation
less clear: the E ARLY heuristic easily beat the R ANDOM            and vice-versa. Another avenue looks at evaluating design
baseline and received reasonably good ratings, which                trade-offs of our approach. We briefly discussed an alter-
suggests it often works well and is not a weak baseline.            native to our photo placement strategy that selects action-
      Therefore, beating the E ARLY baseline demonstrates           photo pairs using ILP. Arguably, each strategy represents a
the effectiveness of our approach. This suggests that there is      different trade-off point between generative flexibility and
merit to reasoning over the structural relationships between        local coherence of the discourse materials. Exploring each
fabula and discourse materials beyond merely how fabula             trade-off point’s expressive range (Smith and Whitehead
causally enables discourse. Rather, multiple aspects of             2010) could be insightful.
concord between fabula and discourse can play a role in                   Reasoning independently about fabula and discourse
the perceived quality of the generated artifact. Note that we       has proven useful for computationally generating interesting
do not claim the weights in Table 1 are optimal, but rather         stories. We believe that exploring their interdependencies
that there exists a configuration of these features that capture    will bring AI story generation systems to new heights.
important relationships between fabula and discourse.
   3
     The Bonferroni correction does not apply as we need to
                                                                                       Acknowledgments
reject at least two out of three null hypotheses to show our main   We thank Darrin Bentivegna for letting us use his photos,
hypothesis.                                                         Kyna McIntosh and Moshe Mahler for the artwork.
References                                  of the 19th ACM International Conference on Multimedia, 223–
                                                                     232.
Aylett, R.; Dias, J.; and Paiva, A. 2006. An affectively
driven planner for synthetic characters. In Proceedings of           Radiano, O.; Graber, Y.; Mahler, M. B.; Sigal, L.; and Shamir,
the 16th International Conference on Automated Planning and          A. in revision. Story albums: Creating fictional stories from
Scheduling, 2–10.                                                    personal photograph sets. Computer Graphics Forum.
Barot, C.; Potts, C. M.; and Young, R. M. 2015. A                    Riedl, M. O. 2009. Incorporating authorial intent into
tripartite plan-based model of narrative for narrative discourse     generative narrative systems. In Proceedings of the AAAI
generation. In Proceedings of the Joint Workshop on Intelligent      Spring Symposium on Intelligent Narrative Technologies II.
Narrative Technologies and Social Believability in Games             Robertson, J., and Young, R. M. 2014. Finding schrödinger’s
at the 11th AAAI Conference on Artificial Intelligence and           gun. In Proceedings of the 10th AAAI Conference on Artificial
Interactive Digital Entertainment, 2–8.                              Intelligence and Interactive Digital Entertainment, 153–159.
Blum, A. L., and Furst, M. L. 1997. Fast planning through            Ronfard, R., and Szilas, N. 2014. Where story and media meet:
planning graph analysis. Artificial Intelligence 90(1):281–300.      computer generation of narrative discourse. In Proceedings of
Dai, Q.; Carr, P.; Sigal, L.; and Hoiem, D. 2015. Family             the 5th Workshop on Computational Models of Narrative, 164–
member identification from photo collections. In Proceedings         176.
of the IEEE Winter Conference on Applications of Computer            Smith, G., and Whitehead, J. 2010. Analyzing the expressive
Vision, 982–989.                                                     range of a level generator. In Proceedings of the 2010
                                                                     Workshop on Procedural Content Generation in Games at the
Domshlak, C., and Mirkis, V. 2015. Deterministic over-
                                                                     5th International Conference on the Foundations of Digital
subscription planning as heuristic search: abstractions and
                                                                     Games, 4–10. ACM.
reformulations. Journal of Artificial Intelligence Research
52:97–169.                                                           Smith, D. E. 2004. Choosing Objectives in Over-Subscription
                                                                     Planning. In Proceedings of the 14th International Conference
Garcı́a-Olaya, A.; de la Rosa, T.; and Borrajo, D. 2011. Using
                                                                     on Automated Planning and Scheduling, 393–401.
the relaxed plan heuristic to select goals in oversubscription
planning problems. In Advances in Artificial Intelligence.           Teutenberg, J., and Porteous, J. 2015. Incorporating global
Springer Berlin Heidelberg. 83–85.                                   and local knowledge in intentional narrative planning. In Pro-
                                                                     ceedings of the 2015 International Conference on Autonomous
Gervás, P. 2009. Computational approaches to storytelling and       Agents and Multiagent Systems, 1539–1546.
creativity. AI Magazine 30(3):49.
                                                                     Tomai, E. 2014. Exploring abductive event binding for
Joshi, D.; Datta, R.; Fedorovskaya, E.; Luong, Q.-T.; Wang,          opportunistic storytelling. In Proceedings of the 10th AAAI
J. Z.; Li, J.; and Luo, J. 2011. Aesthetics and emotions in          Conference on Artificial Intelligence and Interactive Digital
images. IEEE Signal Processing Magazine 28.5:94–115.                 Entertainment, 181–187.
Krishna, R.; Zhu, Y.; Groth, O.; Johnson, J.; Hata, K.; Kravitz,     Ware, S. G., and Young, R. M. 2014. Glaive: A state-
J.; Chen, S.; Kalantidis, Y.; Li, L.-J.; Shamma, D. A.; Bernstein,   space narrative planner supporting intentionality and conflict.
M.; and Fei-Fei, L. 2016. Visual genome: Connecting language         In Proceedings of the 10th AAAI Conference on Artificial
and vision using crowdsourced dense image annotations. In            Intelligence and Interactive Digital Entertainment, 80–86.
arXiv preprint arxiv:1602.07332.
                                                                     Young, R. M.; Ware, S.; Cassell, B.; and Robertson, J. 2013.
Li, B. 2015. Learning Knowledge to Support Domain-                   Plans and Planning in Narrative Generation: A Review of Plan-
Independent Narrative Intelligence. Ph.D. dissertation, Georgia      Based Approaches to the Generation of Story, Discourse, and
Insitute of Technology.                                              Interactivity in Narratives. Sprache und Datenverarbeitung:
Miller, C. E.; Tucker, A. W.; and Zemlin, R. A. 1960. Integer        International Journal of Language Processing 37(1–2):67–77.
programming formulation of traveling salesman problems.              Young, R. M. 2007. Story and discourse: A bipartite model
Journal of the ACM 7(4):326–329.                                     of narrative generation in virtual worlds. Interaction Studies
Montfort, N.; Pérez, R.; Harrell, D. F.; and Campana, A. 2013.      8(2):177–208.
Slant: A blackboard system to generate plot, figuration, and
narrative discourse aspects of stories. In Proceedings of the 4th
International Conference on Computational Creativity, 168–
175.
Morgens, S.-M., and Jhala, A. 2013. Synthetic photographs
for learning aesthetic preferences. In Proceedings of the
Workshop on Late-Breaking Developments in the Field of
Artificial Intelligence at the 27th AAAI Conference on Artificial
Intelligence, 83–85.
Page, E. B. 1963. Ordered Hypotheses for Multiple Treatments:
A Significance Test for Linear Ranks. Journal of the American
Statistical Association 58(301).
Piacenza, A.; Guerrini, F.; Adami, N.; Leonardi, R.; Porteous,
J.; Teutenberg, J.; and Cavazza, M. 2011. Generating story
variants with constrained video recombination. In Proceedings
You can also read